The US Treasury has published a sector-specific framework for financial institutions trying to govern artificial intelligence without treating every new model as an isolated risk project.
Developed through collaboration among more than 100 financial institutions and industry organisations, with input from regulators and technical bodies, the Financial Services AI Risk Management Framework extends the broader NIST approach. Its guidebook translates principles into controls, evidence and maturity stages suited to banking, insurance and other regulated financial work.
The framework’s value is not a promise that AI becomes safe. It is a common structure for deciding which risks matter at each stage of adoption and proving that an institution is managing them.
Maturity determines the control burden
The guidebook begins with an adoption-stage questionnaire. It examines business impact, governance, deployment methods, third-party providers, organisational goals and the sensitivity of data in use.
Institutions are then placed in one of four stages. “Initial” organisations have little or no operational AI. “Minimal” adoption covers limited, low-risk or isolated use. “Evolving” institutions run more complex systems involving sensitive data or external services. At the “embedded” stage, AI plays a significant role in business operations and decisions.
This graduated model avoids two unhelpful extremes. An early-stage firm does not need every possible control before a low-risk experiment. A bank using AI in core decisions cannot rely on the informal oversight appropriate to a pilot.
As systems gain access to sensitive data, customer-facing roles or consequential decisions, the framework adds expectations. Governance grows alongside operational reach rather than arriving after deployment has already become difficult to unwind.
Two hundred and thirty controls create a common language
The framework contains 230 control objectives organised around four functions adapted from NIST: govern, map, measure and manage. A risk-and-control matrix links objectives to adoption stages, while a separate reference guide offers examples of controls and the evidence an institution might retain.
Topics include data quality, fairness and bias, cybersecurity, decision transparency and operational resilience. The framework also recommends AI-specific incident response procedures and a central repository for tracking failures and lessons.
That repository matters because AI behaviour varies with context. Traditional software is generally expected to produce deterministic results from defined inputs, while large language models can respond differently and behave in ways that are harder to predict. Incident records allow institutions to see patterns across systems instead of treating each failure as unrelated.
Trustworthy AI is defined through reliability, safety, security, resilience, accountability, transparency, explainability, privacy and fairness. These qualities extend across the lifecycle. A model that performed well in testing still needs protection from cyber threats and monitoring for changes after launch.
Governance must connect existing risk functions
Financial institutions already operate mature governance, risk and compliance programmes. The Treasury materials do not ask them to replace those structures. They connect AI oversight to them.
Technology teams understand models and infrastructure. Business units know the workflow and consequences of errors. Compliance and risk specialists interpret regulatory duties. Effective oversight requires all of them, because no single group can evaluate the complete system.
The guidebook also leaves room for institutional judgment. It provides control objectives and evidence examples, but each firm must decide which implementation fits its technology, risk appetite and regulatory obligations.
That flexibility is necessary, yet it creates responsibility. A framework can become a checklist if leaders focus on producing documents rather than testing whether controls work. AI incident exercises, monitoring thresholds and clear authority to stop a system are stronger evidence than policy language alone.
The strategic message is straightforward: AI adoption and risk maturity need to advance together. Firms that build common governance can move beyond pilots with greater confidence, while those that add powerful models without strengthening oversight expose themselves to operational, regulatory and reputational damage.
The framework gives financial institutions a map, not an automatic route. Its most important contribution may be a shared vocabulary through which executives, engineers and risk teams can discuss an AI system before its behaviour becomes a business event.
